A large-scale analysis of public-facing, community-built chatbots on Character.AI

TL;DR

This work presents the first large-scale, data-driven examination of public-facing Character.AI greetings, analyzing 2.1 million prompts from about 1 million users to map fandom representation, tropes, and gender-power dynamics in user–bot interactions. Using entity co-occurrence networks, backbone filtering, and Leiden clustering, it identifies 266 fandoms across 4,660 entities, and applies BERTopic to uncover motifs such as toxic relationships, identity exploration, and mental health support. Dependency-parsed analyses reveal that interacting users are depicted as less powerful and more feminine within greetings, highlighting embedded gender norms and the participatory culture surrounding AI-assisted fanfiction. The findings illuminate a new form of online (para)social interaction at the intersection of generative AI and user-generated content, while emphasizing ethical considerations, potential risks to youth, and the need for open data to support further research.

Abstract

This paper presents the first large-scale analysis of public-facing chatbots on Character.AI, a rapidly growing social media platform where users create and interact with chatbots. Character.AI is distinctive in that it merges generative AI with user-generated content, enabling users to build bots-often modeled after fictional or public personas-for others to engage with. It is also popular, with over 20 million monthly active users, and impactful, with recent headlines detailing significant issues with youth engagement on the site. Character.AI is thus of interest to study both substantively and conceptually. To this end, we present a descriptive overview of the site using a dataset of 2.1 million English-language prompts (or ``greetings'') for chatbots on the site, created by around 1 million users. Our work explores the prevalence of different fandoms on the site, broader tropes that persist across fandoms, and how dynamics of power intersect with gender within greetings. Overall, our findings illuminate an emerging form of online (para)social interaction that toes a unique and important intersection between generative AI and user-generated content.

Figures (8)

• Figure 1: The page of a popular chatbot on character.AI. The character has been created by a user of the site for interaction by themselves and others. Text in the dark grey box represents the character's greeting; these greetings are the main focus of analysis in the present work.
• Figure 2: a) The proportion of all following relationships (y-axis) accounted for by the top X% of users (x-axis). b) The proportion of all interactions (chats; y-axis) accounted for by the top X% of characters (x-axis)
• Figure 3: Proportion of all interactions (unique chats) with any character containing at least one named entity related to a fandom (y-axis) for the 14 different high-level types of fandoms we identified
• Figure 4: Proportion of all interactions (unique chats) with any character containing at least one named entity related to specific fandoms (y-axis) for the top 15 fandoms, colored by fandom type
• Figure 5: Top 15 topics (y-axis) by proportion of all interactions with characters in our dataset (x-axis)